What should force an agent loop to stop before it tries again?

[Keesan12]

We kept seeing the same failure pattern in autonomous coding runs:

not one dramatic bug, but a loop that quietly kept going after the useful part was already over.

That usually looked like some mix of:

• retrying without new evidence
• tests that passed without really increasing confidence
• context getting larger while the real failure got less clear
• overnight runs spending more than the work was worth

That is a big part of why we built MartinLoop.

The thing we kept wanting was not "more autonomy." It was a better stop rule and a cleaner receipt:

• what failed
• what changed
• what passed
• why the next attempt is actually different

Curious how other people here think about this:

What should force an agent loop to stop before it tries again?

A few candidates we've run into:

• no new evidence since the last attempt
• cost crossed a threshold relative to task value
• tests passed, but only generated tests passed
• the changed-file surface keeps growing without narrowing the problem
• the run can no longer explain why the next attempt is meaningfully different

If you're running agents overnight or in longer loops, what do you actually treat as the stop condition?